A-242 Integrating Respiratory Metagenomics and Metatranscriptomics for Diagnosis of Lung Cancer and Infection in Patients with Pulmonary Diseases

Abstract

Abstract Background Advances in unbiased metagenomic next generation sequencing (mNGS) technologies have enabled the study of microbial and host genetic material (DNA and RNA) in one test. In this study, we aimed to develop machine learning-based differential diagnostic models (MLBDDMs) using the metagenomic and human transcriptomic data generated by an affordable bronchoalveolar lavage fluid (BLAF) mNGS assay and investigated their clinical utility for early differential diagnosis of lung cancer and pulmonary infection in patients with pulmonary diseases. Methods We recruited 775 patients with respiratory disease, including 160 pathologically diagnosed lung cancer and clinically diagnosed 615 infectious causes (131 tuberculosis, 172 fungal pneumonia and 312 bacterial pneumonia). An affordable mNGS assay on BALF samples collected from these patients on admission were performed. Using the generated mNGS data, we compared the differences in microbial diversity and host gene expression between lung cancer patients and pulmonary infection patients. The BLAF mNGS datasets of lung cancer group and each infection group were then randomly divided into a training dataset and a validation dataset at a ratio of approximately 3:1 for developing optimal MLBDDMs that can be used to distinguish lung cancer from various pulmonary infections. Results By comparing the BALF mNGS data of lung cancer (n = 160) and pulmonary infection (n = 615), we found that the infection group had higher microbial diversity than lung cancer group (P-value < 0.05). Respiratory colonizing microorganisms (e.g., Corynebacterium propinquum and Bacteroides uniformis) and pathogen (Mycobacterium tuberculosis and Cryptococcus neoformans) were found as differential microbes (adjusted p-value < 0.05, LDA score > 2). From BALF gene expression data, we detected 175 genes enriched in NOD-like receptor signaling pathway and chemokine signaling pathway differentially expressed between lung cancer and pulmonary infection groups (False Discovery Rate, FDR < 0.05). Cell composition analysis revealed that macrophage M1 was higher in lung infection group (P-value < 0.001), whereas mast cell activated and DCs activated were higher in lung cancer group (P-value < 0.001, P-value = 0.016). We integrated the metagenomic (microbial composition and human copy number variation) and transcriptomic data (host differentially expressed genes and cell composition) generated by the BALF mNGS assay with eleven machine learning classifiers to establish diagnosis models for distinguishing lung cancer from pulmonary infection (we named LC/PI model). The results showed that a Random Forest diagnostic model (the RF-LC/PI model) had optimal performance, with a sensitivity and specificity of 86.7% and 87.8%, respectively, in distinguishing lung cancer from pulmonary infection (area under the receiver operating characteristic curve [AUC] = 0.838 in the training dataset; AUC = 0.79 in a held-out validation dataset). Similar to the establishment of the LC/PI model, we further developed three diagnostic models for distinguishing lung cancer and tuberculosis (LC/TB model), lung cancer and fungal pneumonia (LC/FP model), and lung cancer and bacterial pneumonia (LC/BP model), respectively. The AUC of these three models were 0.91, 0.88, 0.91, respectively, showing a high differential diagnosis accuracy. Conclusions We have established MLBDDMs using BALF metagenomic and metatranscriptomic data and achieved superior accuracy for differentiating lung cancer and pulmonary infections, which could promote early diagnosis of pulmonary diseases and benefit more patients with one test.